As we delve into ASI, this will be our most philosophical blog to date, but we also want to ground the conclusions in practical knowledge. We discuss what might be possible in the future a long time from now, and how this will affect current development. First, we must try to define what type “intelligence” we are trying to emulate. This is important because if we decide on the definition of “intelligence” then we can create an North Star for both research and practical applications and helps us understand if we are getting us closer to creating our idealized form of “AI”.
We want to define “intelligence” very narrowly. We will confine our definition to goal-oriented behaviour. This is to avoid confusing the efficiency of a machine for what we value has human beings. We value people for how they make us feel as much as for their ability to solve a problem. The subjective experience of an enjoyable meal with friends cannot be easily deconstructed into a goal-oriented task but is a crucial to our sense of worth has humans. So the narrow definition of intelligence becomes crucially important when we think of what we are designing to be “intelligent”.
If you assume that a super intelligent computer can be designed by making a replica of the human brain, then it is very possible that this “brain” will not be human. What do we mean by “not-human”? Descartes, the famous dualist held that non-human animals, and by extension, machines, were entirely physical automata systems – they are reducible to simple machines that respond in predictable ways to stimuli. More recently David Chalmers has discussed conceptual zombies; creatures that resemble us in every physical sense but do not have conscious experiences. That is to say, bio-machines that are not human. And some scientists today are confirming that some of our mental states are controlled by our own gut bacteria meaning some of our own subjective experiences (hunger for example) may be under the control of colonies of microbes. If we accept dualism or related concepts, then we are free to build machines that are very good at specific tasks. Free because if a constructed machine does not have conscious experiences, then it cannot be human and not endowed with human values and the machine can be made to do any task however demeaning it might be to a human.
How would we go about building such a system? And what would this mean for development goals? If you believe that the brain is just a collection of elements that self-organize into dynamic structures that process stimuli, then we are well on our way to some interesting new discoveries. Through the increasing sophistication of neural networks, the processing speed and complexity of sub-systems, we are on our way to creating very flexible systems – a machine that both plays Go but also uses that pattern recognition “intelligence” to find our lost keys. These machines will not be self-directed because of our assumption of dualism, but they will be very flexible and will be able to handle a wide range of tasks, You will see further development with increasingly more complicated tasks being handled by newer networks as we progress our research. In this case, companies that succeed will be developing physical system replicas.
However, some do not believe the brain is a collection of elements. They do not believe that the function of a brain is reducible to simple collections of individual parts. They believe the brain really is a about information processing architecture manipulating symbolic representations, and not the underlying elements upon which the processing is achieved. A Lego set may allow you to build a replica of a person, but it has none of the software that makes a person run. If this interpretation is more to your taste, then much of the current development might be misguided. And you might be better suited following and investing your time in creating software that can more quickly handle and process information by combing inputs across systems and synthesising them in a decision matrix perhaps.
These two approaches are not mutually exclusive and they each have something to offer the other. You can develop better networks that allow information to be stored and retrieved (attention networks) and you can run your networks on distributed systems with software processing the interactions between them. This two-paragraph description is a very simple summary and we have not even touched upon the questions of valid statistical representations of reality. But if you are interested in more, please contact us via our website.
At Delvify we try to avoid arguments about what might be and focus more on the practical approaches. We leave the discussions of alternate simulated universes to fiction writers. By concentrate on the practical applications of research – what can we achieve given our understanding of how intelligence in formed, we can choose the best applications of research.